CN110866799A

CN110866799A - System and method for monitoring online retail platform using artificial intelligence

Info

Publication number: CN110866799A
Application number: CN201910801326.7A
Authority: CN
Inventors: 陈力; 刘石竹; 黄凯琳; 杨尚林; 周辉
Original assignee: Beijing Jingdong Shangke Information Technology Co Ltd; JD com American Technologies Corp
Current assignee: Beijing Jingdong Shangke Information Technology Co Ltd; JD com American Technologies Corp
Priority date: 2018-08-28
Filing date: 2019-08-28
Publication date: 2020-03-06
Anticipated expiration: 2039-08-28
Also published as: CN110866799B; US10853697B2; US20200074242A1

Abstract

A method and system for monitoring an e-commerce platform. The method comprises the following steps: receiving, by a computing device, feedback submitted by a user through an e-commerce platform; generating, using an AI processor, a vector based on the content of the feedback, the context of the feedback, and the user profile; the quantities are classified using an AI classifier to determine the function and the state of the function corresponding to the feedback. The content comprises characters, voice, images and videos; the context includes time, location and submission channel of the feedback; the profile includes attributes, history and preferences of the user. The dimensions of the vector correspond to text, speech, image, video, time, location, submission channel, attributes, history and preferences of the user, respectively.

Description

System and method for monitoring online retail platform using artificial intelligence

Cross Reference to Related Applications

Some references, which may include patents, patent applications, and various publications, are cited and discussed in the description of the present disclosure. Citation and/or discussion of such references is provided merely to clarify the description of the present disclosure and is not an admission that any of these references are "prior art" with respect to the disclosure described herein. All references cited and discussed in this specification are herein incorporated by reference in their entirety to the same extent as if each reference were individually incorporated by reference.

Technical Field

The present disclosure relates generally to monitoring the health status of an e-commerce platform, and more particularly, to a system and method for monitoring the health of an online retail platform in real-time through deep learning based on feedback from a user.

Background

The background description provided herein is for the purpose of generally presenting the context of the disclosure. Work of the presently named inventors, to the extent it is described in this background section, as well as aspects of the description that may not otherwise qualify as prior art at the time of filing, are neither expressly nor impliedly admitted as prior art against the present disclosure.

Electronic commerce has experienced a high rate of growth over the years. Huge online retail platforms such as amazon, arbiba and kyoto have hundreds of millions of active users with total revenue in the billions. In addition, both revenue and user numbers are growing rapidly (40% quarterly in 2017 for second quarterly traffic in the case of the kyoto).

On the one hand, the huge traffic is accompanied by a large usage of online retail platforms (mobile applications and websites). The heavy use presents challenges to the usability and stability of the platform. Thus, unknown platform failures will pose a hazard to user experience, revenue and public reputation, leading to serious financial and social consequences.

On the other hand, the high-speed growth is attributed to the rapid increase in website functionality and/or features. On an online retail website, there are hundreds of channels of commodity products and a range of functions including search, recommendation, shopping cart, delivery, and payment. All of these functions and channels are developed or owned by different teams within the company. These functions typically overlap each other, and more often either the upstream or downstream portions depend on each other. This complexity poses obstacles to each team accurately understanding the health of the functions they own, diagnosing errors, and providing solutions.

Accordingly, there is an unresolved need in the art to address the above-mentioned deficiencies and inadequacies.

Disclosure of Invention

In certain aspects, the present disclosure relates to a method for monitoring the health of an e-commerce platform. In certain embodiments, the method comprises: receiving, by a computing device, feedback submitted by a user through an e-commerce platform; generating, by a feedback processor of a computing device, a vector based on content of the feedback, context of the feedback, and a profile of the user; and classifying, by a classifier of the computing device, the quantities to obtain a function of the e-commerce platform and a status of the function corresponding to the feedback, and to prepare an alert when the status is a fault. The content includes at least one of text, voice, image, and video; the context includes at least one of a time to submit the feedback, a location to submit the feedback, and a submission channel of the feedback; the user profile includes at least one of attributes of the user, a purchase history of the user, and preferences of the user using the e-commerce platform. The vector has a predetermined number of dimensions, and each of the text, speech, image, video, time at which the feedback was submitted, location at which the feedback was submitted, channel at which the feedback was submitted, attributes of the user, purchase history of the user, and purchase preferences of the user corresponds to at least one dimension of the vector.

In certain embodiments, the feedback processor and the classifier are performed using at least one artificial intelligence model.

In some embodiments, the step in which the vector is generated comprises: the content is processed using a feedback processor to obtain content dimensions for vectors corresponding to text, speech, images, and video. In certain embodiments, the method further comprises: the content is cleaned before processing the content to obtain the content dimensions of the vector. In certain embodiments, the method further comprises: the method includes the steps of separating an image into an image text and a background image, processing the image text to obtain an image text result and processing the background image to obtain a background image result, and integrating the image text result and the background image result to obtain a content dimension of a vector corresponding to the image.

In certain embodiments, the method further comprises: the method further includes sending the alert to an administrator of the e-commerce platform responsible for the function, receiving an instruction from the administrator corresponding to the alert when the alert is false, and retraining the feedback processor and classifier using the instruction.

In some embodiments, the classifier is trained using a plurality of historical feedbacks and a functional class structure, the functional class structure comprising: including e-commerce platform websites, e-commerce platform applications, and layer 1 categories of e-commerce platform external links. In some embodiments, the level 1 categories of websites include level 2 categories of product pages, shopping carts, and payments; the level 2 categories of the product page include level 3 categories of product description, product search, and product recommendation.

In some embodiments, the classifier includes a plurality of classification models, each classification model providing a candidate function based on each historical feedback, and the candidate functions provided by the classification models are used by the integration model to determine a function corresponding to each feedback.

In certain aspects, the present disclosure relates to a system for monitoring the health of an e-commerce platform. In certain embodiments, a system includes a computing device. The computing device has a processor and a storage device storing computer executable code. When the computer executable code is executed at a processor, it is configured to perform the above-described method.

In certain aspects, the present disclosure relates to non-transitory computer-readable media storing computer-executable code. When the computer executable code is executed at a processor of a computing device, it is configured to perform the method as described above.

These and other aspects of the present disclosure will become apparent from the following description of the preferred embodiments, taken in conjunction with the accompanying drawings and the description thereof, wherein changes and modifications may be made therein without departing from the spirit and scope of the novel concepts of the disclosure.

Drawings

The drawings illustrate one or more embodiments of the disclosure and, together with the written description, serve to explain the principles of the disclosure. Wherever possible, the same reference numbers will be used throughout the drawings to refer to the same or like elements of an embodiment.

FIG. 1 schematically depicts a workflow of an e-commerce platform monitoring system in accordance with certain embodiments of the present disclosure.

FIG. 2 schematically depicts an e-commerce platform monitoring system in accordance with certain embodiments of the present disclosure.

Fig. 3 schematically depicts an AI processor according to certain embodiments of the present disclosure.

Fig. 4 schematically depicts an image processing process according to certain embodiments of the present disclosure.

Fig. 5 schematically depicts feature vectors according to certain embodiments of the present disclosure.

Fig. 6 schematically depicts a feature matrix according to certain embodiments of the present disclosure.

Fig. 7 schematically depicts an AI classifier according to certain embodiments of the present disclosure.

Fig. 8 schematically depicts a database according to certain embodiments of the present disclosure.

FIG. 9 schematically depicts a method for training an e-commerce platform monitoring system, in accordance with certain embodiments of the present disclosure.

Fig. 10 schematically depicts the structure of the functionality according to certain embodiments of the present disclosure.

Fig. 11 schematically depicts an integrated structure according to certain embodiments of the present disclosure.

Fig. 12 schematically depicts a method of integrating all one-to-many (one-summary-all) classifiers in accordance with certain embodiments of the present disclosure.

FIG. 13 schematically depicts a method for using an e-commerce platform monitoring system, in accordance with certain embodiments of the present disclosure.

Detailed Description

The present disclosure is more particularly described in the following examples, which are intended as illustrations only, since numerous modifications and variations therein will be apparent to those skilled in the art. Various embodiments of the present disclosure will now be described in detail. Referring to the drawings, like numbers indicate like components throughout the views. As used in the description herein and throughout the claims that follow, the meaning of "a", "an", and "the" includes plural references unless the context clearly dictates otherwise. Furthermore, as used in the description herein and throughout the claims that follow, the meaning of "in. Also, headings or subheadings may be used in the description for the convenience of the reader without affecting the scope of the disclosure. In addition, some terms used in the present specification are defined more specifically below.

Terms used in this specification generally have their ordinary meanings in the art, in the context of the present disclosure, and in the specific context in which each term is used. Certain terms used to describe the present disclosure are discussed below or elsewhere in the specification to provide additional guidance to the practitioner regarding the description of the present disclosure. It should be understood that the same thing can be said in more than one way. Thus, alternative language and synonyms may be used for any one or more of the terms discussed herein, and have no special meaning as to whether a term is set forth or discussed in detail herein. Synonyms for certain terms are provided. Recitation of one or more synonyms does not exclude the use of other synonyms. The use of examples anywhere in this specification, including examples of any terms discussed herein, is illustrative only and in no way limits the scope and meaning of the disclosure or any exemplary terms. As such, the present disclosure is not limited to the various embodiments presented in this specification.

Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure belongs. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and the present disclosure, and will not be interpreted in an idealized or overly formal sense unless expressly so defined herein.

As used herein, "about," "approximately," or "approximately" generally means within 20%, preferably within 10%, more preferably within 5% of a given value or range. Numerical values set forth herein are approximate, meaning that the term "left or right," "about," "approximately," or "approximately" can be inferred if not expressly stated.

As used herein, "plurality" means two or more.

As used herein, the terms "comprising," "including," "carrying," "having," "containing," "involving," and the like are to be understood as being open-ended, i.e., meaning including but not limited to.

As used herein, at least one of the phrases A, B and C should be construed to mean a logic (a OR B OR C) using a non-exclusive logic "OR" (OR). It should be understood that one or more steps of a method may be performed in a different order (or simultaneously) without altering the principles of the present disclosure. As used herein, the term "and/or" includes any and all combinations of one or more of the associated listed items.

As used herein, the term "module" may refer to a portion of or include: an Application Specific Integrated Circuit (ASIC); an electronic circuit; a combinational logic circuit; a Field Programmable Gate Array (FPGA); a processor (shared, dedicated, or group) that executes code; other suitable hardware components that provide the described functionality; or some or all of the above, for example in a system on a chip. The term module may include memory (shared, dedicated, or group) that stores code executed by the processor.

The term code, as used herein, may include software, firmware, and/or microcode, and may refer to programs, routines, functions, classes, and/or objects. The term shared, as used above, means that some or all code from multiple modules may be executed using a single (shared) processor. In addition, some or all code from multiple modules may be stored by a single (shared) memory. The term group, as used above, means that some or all code from a single module may be executed using a group of processors. In addition, a set of memories may be used to store some or all of the code from a single module.

The term "interface" as used herein generally refers to a communication tool or device at the point of interaction between components for performing data communications between the components. In general, the interface may be adapted to both hardware and software levels, and may be a unidirectional or bidirectional interface. Examples of physical hardware interfaces may include electrical connectors, buses, ports, cables, terminals, and other I/O devices or components. The components in communication with the interface may be, for example, components or peripherals of a computer system.

The present disclosure relates to computer systems. As shown in the figures, computer components may include physical hardware components, shown as solid line blocks, and virtual software components, shown as dashed line blocks. Unless indicated otherwise, those of ordinary skill in the art will appreciate that such computer components may be implemented in, but are not limited to, software, firmware, or hardware components, or a combination thereof.

The apparatus, systems, and methods described herein may be implemented by one or more computer programs executed by one or more processors. The computer program includes processor-executable instructions stored on a non-transitory tangible computer-readable medium. The computer program may also include stored data. Non-limiting examples of the non-transitory tangible computer readable medium are nonvolatile memory, magnetic storage, and optical storage.

The present disclosure now will be described more fully hereinafter with reference to the accompanying drawings, in which embodiments of the disclosure are shown. This disclosure may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

As mentioned above, there is a need to accurately monitor the health of an e-commerce platform. In some embodiments, the monitoring method is to make a platform log and set an index to be monitored, i.e., if the number of clicks increases sharply in a short time, it may be that a malfunction occurs and an alarm is sent. However, the index is also affected by many other factors besides the fault. For example, the increase in the number of clicks may be due to traffic growth or vacation, rather than failure. Therefore, a more direct and accurate source of information is needed to monitor health status.

In some embodiments, the user's feedback is considered an information source. However, the traditional way of manually checking user feedback is not suitable for huge online retail platforms. Platforms have hundreds or even thousands of functions, so it is difficult for people to remember all definitions and nuances, let alone to give an accurate response. Furthermore, the delay is high because one needs time to understand the feedback, check the reference and the response. Finally, the cost of maintaining a large team is high.

In certain aspects, the present disclosure provides a self-contained self-completed (self-defined) system for monitoring the health status of an online retail platform. By utilizing the user's feedback and integrating it with knowledge about the platform, the system can detect functional problems in a timely, accurate, and automated manner via the ability to utilize Artificial Intelligence (AI) including natural language processing, computer vision, and machine learning.

In detail, the system establishes a knowledge base about team structures, application function structures, and correspondence between the two of the electronic commerce companies. At the same time, the system immediately parses the user feedback in various formats (text, voice, image, video) and extracts useful information from it via AI techniques. Finally, the system combines the two sources of information to make a decision-sending an alarm to a specific team to fix the reported problem. The information extracted from the feedback includes: which function is not healthy and which measures should be taken to repair it. The measures comprise the following steps: find a team with this functionality, inform the team that this is problematic, and give a reason and advise action.

In certain embodiments, the present disclosure relates to software systems embedded with AI technology to enable instant, accurate health monitoring of online retail platforms based on user feedback. The various feedback formats require the uniqueness of the AI technique used here-there are a wide variety of inputs (structured data, unstructured data, text, image audio, video), so the system has an intensive integration approach to integrate all the different data formats to create the most accurate problem reports.

FIG. 1 schematically illustrates a global workflow of a software system according to certain embodiments of the present disclosure. As shown in FIG. 1, a user 102 performs an act 104 of submitting feedback 106. The submitted feedback has multimedia content 1060, the multimedia content 1060 including text 1062, audio or speech 1064, images 1066, and video 1068. The system records the context of submission 1040, the context of submission 1040 including the time 1042 at which the feedback was submitted, the location 1044 at which the user submitted the feedback, and the channel 1046 of submission. The submission channel includes an Application (APP) or a website. The submission is related to a user Identification (ID), and the system may access a user profile 1020 based on the identified user ID, the user profile 1020 including attributes 1022, history 1024, and preferences 1026. Attributes 1022 include registration information of the user, such as gender, age, hobbies, email address. The history 1024 includes the user's purchase history and optionally a history of the user's feedback or other actions of the user using the e-commerce platform. Preferences 1026 include user preferences for using the website, such as his products of interest based on his search history. The content data is first processed and dumped as structured data via the AI processor 108. The data pool 112 stores structured data from the AI processor 108, submitted context data 1040, and profile data 1020 of the user, the data pool 112 configured to integrate the structured data from the AI processor 108, the context 1040 related to the submission of the feedback, and the profile 1020 related to the submission of the feedback to form a vector or matrix of one or more feedbacks. In training the system, a matrix is generated based on a number of feedbacks. When the system is used for monitoring, a vector may be generated for each feedback, which is then analyzed to determine its function and health status. After vector or matrix generation or integration, the data pool 112 then sends the vector or matrix to the AI classifier 114. At step 116, the machine learning classifier 114 is applied to predict whether the feedback 106 is related to a function, and if so, whether the function is healthy at step 118, and which function is in question at step 120. The system then matches the responsibility owner or administrator for the function based on the function owner knowledge 132 and sends a fault report or alarm 122 to the responsibility owner 124 for the function. When the function owner determines that the alarm is true at step 126, it repairs the fault at step 128. If the alarm is false, the function owner sends a false alarm to the database 130 and provides the false alarm to the function owner knowledge 132 to retrain the AI classifier 114 using the updated function owner knowledge. The data pool 112 and database 120 may be an integrally formed database that includes the context of historical feedback 1040, the platform user's profile 1020, matrices and vectors generated by the AI processor 108, functional labels of the feedback (either manually added or generated during AI training), and the like.

During this process, AI techniques are used to process the original content and make decisions. For AI, data acquisition is essential for the success of the AI model. In certain embodiments, the system stores a large amount of historical data in a database to train the AI model. In addition, an online training mechanism is enabled — once the AI is in error (e.g., a false alarm is sent), an error message will be sent immediately back to the AI model for online retraining. Thus, the system is self-completing in real time.

FIG. 2 schematically depicts a health monitoring system for an e-commerce platform, in accordance with certain embodiments of the present disclosure. The system 200 shown in fig. 2 and the system 100 shown in fig. 1 are different ways of showing the same or similar systems. As shown in fig. 2, system 200 includes a server computing device 210, a plurality of management computing devices 250, and a network 240 connecting management computing devices 250 with server computing device 210.

The server computing device 210 may act as a server or a host computer. In certain embodiments, the server computing device 210 may be a cloud computer, server computer, cluster, general purpose computer, or special purpose computer that provides platform monitoring services. In certain embodiments, the management computing device 250 may be a cloud computer, a mobile device, a tablet, a general-purpose computer, a headless computer, a wearable device, or a special-purpose computer that receives alerts from the server computing device 210 and, in response to the alerts, sends an assessment of the alerts to the server computing device 210. In some embodiments, the network 240 may be a wired or wireless network and may be in various forms, such as a public network and a private network. Examples of a network may include, but are not limited to, a LAN or a Wide Area Network (WAN) including the Internet. In some embodiments, two or more different networks and/or interfaces may be employed to connect server computing device 210 to user computing device 250. In some embodiments, interface 240 may also be a system interface, a Universal Serial Bus (USB) interface.

As shown in fig. 2, server computing device 210 may include, but is not limited to, a processor 212, a memory 214, and a storage device 216. In some embodiments, the server computing device 210 may include other hardware components and software components (not shown) to perform their corresponding tasks. Examples of such hardware and software components may include, but are not limited to, other desired memories, interfaces, buses, input/output (I/O) modules or devices, network interfaces, and peripherals.

Processor 212 may be a Central Processing Unit (CPU) configured to control the operation of server computing device 210. The processor 212 may execute an Operating System (OS) or other application of the server computing device 210. In some embodiments, the server computing device 210 may have more than one CPU as a processor, such as two CPUs, four CPUs, eight CPUs, or any suitable number of CPUs. The memory 214 may be volatile memory, such as Random Access Memory (RAM), for storing data and information during operation of the server computing device 210. In some embodiments, memory 214 may be a volatile memory array. In some embodiments, the server computing device 210 may run on more than one memory 214. The storage device 216 is a non-volatile data storage medium for storing an OS (not shown) and other applications of the server computing device 210. Examples of storage device 216 may include non-volatile memory, such as flash memory, memory cards, USB drives, hard drives, floppy disks, optical drives, Solid State Drives (SSDs), or any other type of data storage device. In certain embodiments, the storage device 216 may be local storage, remote storage, or cloud storage. In some embodiments, the server computing device 210 may have multiple storage devices 216, which may be the same storage device or different types of storage devices, and applications of the server computing device 210 may be stored in one or more storage devices 216 of the computing device 210. As shown in FIG. 2, the storage device 216 includes a platform monitor 220. The platform monitor 220 uses feedback from the user or customer to provide a service that monitors the e-commerce platform.

Among other things, the platform monitor 220 includes a feedback extraction module 222, an AI processor 224, a vector generator 226, an AI classifier 228, an administrator communication module 230, and a database 232. In some embodiments, platform monitor 220 may include other applications or modules as needed for the operation of modules 222-232. It should be noted that the modules are each implemented by computer-executable code or instructions or data tables or databases, which together form an application. In some embodiments, each module may also include a sub-module. Alternatively, some modules may be combined into one stack. In other embodiments, some modules may be implemented as circuitry rather than executable code. In some embodiments, some or all of the modules of the platform monitor 220 may be located at a remote computing device or cloud device.

The feedback extraction module 222 is configured to retrieve or receive feedback submitted by a user of the e-commerce platform, extract content from the feedback, and send the extracted content to the AI processor 224. The feedback content includes at least one of text, audio or speech, images and video.

In certain embodiments, during training of the platform monitor 220, the feedback extraction module 222 is configured to retrieve and extract historical feedback stored in the database 232. To ensure efficient training of the platform monitor 220, the feedback extraction module 222 may only provide high quality historical user feedback. These raw text, images, audio, video are collected from daily operations, all of which are stored in an internal database that can be used to train the AI algorithm.

In certain embodiments, during operation of the platform monitor 220, new feedback is added to the database 232, the platform monitor 220 is configured to check the database 232 at short predetermined time intervals, and batch process the newly added feedback.

In certain embodiments, the platform 220 examines the feedback in real-time, and the feedback extraction module 222 is configured to receive the feedback one at a time and send the content of the extracted one to the AI processor 224.

The AI processor 224 is configured to convert the content into structured content data, such as a content vector or a content matrix, upon receiving the content from the feedback extraction module 222. Referring to fig. 3, the AI processor 224 includes a content pre-processor 2240, a text processing module 2242, a voice recognition module 2244, an image processing module 2246, and a video processing module 2248.

The content pre-processor 2240 is configured to remove noise from the content to provide the cleaned data to the text processing module 2242, the voice recognition module 2244, the image processing module 2246, and the video processing module 2248. The text processing module 2242 is configured to convert the text into a numerical value upon receiving the cleaned text. The speech recognition module 2244 is configured to recognize text from the audio and convert the recognized text into a numerical value upon receiving the cleaned audio. The image processing module 2246 is configured to, upon receiving the cleaned image, separate the image into text and background images (the image portion with the text is removed from the image), process the text and background images separately, and integrate the results into a numerical value. The video processing module 2248 is configured to, upon receiving the washed video, separate the image of the video into text and background images, process the text and background images separately, and integrate them to obtain a numerical value. In processing one feedback, those values from the processing content are defined as dimensions of the content vector, and in processing multiple feedbacks (e.g., during training), those values from the processing content are defined as dimensions of the content matrix.

In some embodiments, the text processing module 2242 is configured to obtain a feature or value of the text upon receiving the scrubbed text from the content processor 2240. In particular, text processing module 2242 first partitions each text into word sequences and then characterizes the words into dimensions of content vectors, representing word occurrences, word co-occurrences, part of speech, named entities, sentence syntactic structure, and word senses [8 ]. Many techniques may be used: for example, n-grams and tfidf are used to represent word occurrences [8], word2vec [9] is used to represent the context of a word (co-occurring with other words) [8], POS and named entities identify parts of speech and named entities used to find words, and further syntactic and semantic analysis is applied to obtain the syntactic role and semantics of the word [8 ]. During training, the characterization results in a large matrix with columns as the features and rows as feedback. The processing of the plurality of texts is used to train the model or use the model. During operation, when applying a batch process with new feedback, characterization of the batch feedback also results in a matrix. During operation, as applications process new feedback one by one, text processing module 2242 may also process one feedback to produce a dimension of a vector.

In some embodiments, the image processing module 2246 is configured to obtain features or values of the image upon receipt of the image from the content pre-processor 2240 (or from the feedback extraction module 222 if data cleansing is not required at this stage). Specifically, the image processing module 2246 separates the image into a text (image text) and a background image extracted from the image, processes the image text and the background image separately, and integrates the two to obtain a numerical value. In certain embodiments, image processing module 2246 applies convolutional neural network [10] or deep neural network based techniques to represent the background image as a meaningful content vector dimension.

Fig. 4 schematically shows an example of obtaining content vector dimensions of an image. As shown in fig. 4, a screenshot 402 of the phone screen is provided in the feedback. First, the image processing module 2246 separates the original image 402 into two parts, i.e., the image text of "what you are looking for" and the background image 404 with the image text removed. Extraction of image text may be performed using OCR methods trained on internal data. And extracting the text box by using the extracted text box and the coordinates thereof in the image, and obtaining a background image. Then image processing module 2246 processes the image text to find what the keywords "find" and syntax structure "are. "in some embodiments, the keywords are selected by tfidf + ngram, and the syntactic structure is selected by matching the image text to a predefined domain-specific structure. In some embodiments, all image text extracted from images in the training data is collected in order to predefine sentence structures. The image text is then clipped into sentences. Each pair of sentences is obtained, and a similarity score between the two sentences is calculated according to the words in each pair of sentences. Specifically, each sentence is taken as a set of words S ═ w, and the similarity between sentence i and sentence j is defined as:

s(i,j)＝|S_i∩S_j|

using the similarity score, the sentences in the training data are divided into M groups using a clustering method K-means: g₁,…,G_M. M is predefined based on the number of syntactic structures estimated in the text corpus. Each group has sentences with similar word vectors. For the new sentence k, its nearest group is calculated using the following formula:

G(k)＝argmin_G1,…,GM(∑_i∈Gms(k,i))

following a similar process, background images are rendered into vectors based on their image representation via, for example, the AutoEncoder method. The similarity between the background images is then defined as cosine similarity, and the background images are grouped. For a new image, its closest set of images can be obtained. As shown in fig. 4, the likelihood of a group is obtained based on image text or background images, and the two results can be integrated to obtain a more accurate estimate of the group or related function of the original image. In some embodiments, the group ID of the OCR text and background image is used as an additional feature, concatenated after the text dimension of the content vector or content matrix. As shown in the results in fig. 4, the probability of the search page being 96% is given according to the execution result of the image processing module 2246, which may form the correspondence dimension value 1 of the content vector or the content matrix corresponding to the group ID of "search page".

The audio may be identified to obtain text and the video may be separated into images, the processing of the audio and video being similar to the process described above with respect to text and images.

The vector generator 226 is configured to, upon receiving the content vector or content matrix from the AI processor 224, retrieve the values of the context and the values of the user profile, append these values to the content vector or content matrix to form a vector or matrix. In certain embodiments, the context and profile are stored in database 232. In some embodiments, if the values of the context and user profile are not available in the database 232, the vector generator 226 is further configured to convert the context and user profile into values. In some embodiments, the conversion is performed using virtual variables. For example, if the submitted location 1044 has 100 cities, 100 virtual variables are provided to represent them separately. If the city is "Beijing" (e.g., the first city), the first virtual variable is set to 1 and the other virtual variables are set to 0. In certain embodiments, the vector generator 226 may also be part of the AI processor 224.

Fig. 5 schematically illustrates a vector or (or named feature vector) of feedback generated by the vector generator 226, according to some embodiments of the present disclosure. As shown in fig. 5, the vector includes features from text, images, context of submission, and user profile. The image features include image text features and background image features. In some embodiments, the vector includes a predetermined number of dimensions. As shown in fig. 5, the vector includes m + n + p + q +8 dimensions, where each of m, n, p, and q is a positive integer. In some embodiments, the vector includes approximately 5,000 dimensions, and most of the dimensions correspond to words or phrases from text 1062. The characteristic dimensions shown in fig. 5 are for illustration only and may vary during operation. For example, time may be divided into dimensions of seasons, months, weekdays, hours of a day, and virtual variables are used to define each dimension. In some embodiments, each of the location, submission channel, and attributes are defined by a virtual variable. In some embodiments, the history of the user refers to a history of complaints and the activity of the user using the website, the history being defined by one or more real numbers. After generating the vector, the feature vector generator 226 sends the vector to the AI classifier 228 to make decisions-notifying the fault.

During training or batch processing, the vector generator 226 is configured to generate a matrix (or named feature matrix) instead of a vector. Fig. 6 schematically illustrates a feedback matrix according to certain embodiments of the present disclosure. As shown in fig. 6, each row of the matrix is a vector corresponding to one feedback. Each row includes text features (word features, syntactic features, semantic features), image text features, background image features, context features, and user profile features, which are represented by values, and the values in a row are the dimensions of the vector corresponding to the feedback. In some embodiments, the feedback in a row may not have a corresponding characteristic for each dimension, assigning the dimension that is missing from the feedback a value of 0. Note that each row of the matrix shown in fig. 6 includes similar information to the vector shown in fig. 5.

The AI classifier 228 is configured to determine, upon receiving the vector/matrix from the vector generator 226, whether the feedback relates to a function of the e-commerce platform, a status of the function, and which function. Referring to fig. 7, the AI classifier 228 includes a function determination module 2280, a function status module 2282, and a reporting module 2284. The function determination module 2280 is configured to process the vector/matrix using various classification models to obtain one or more functions related to the vector (or each row of the matrix), and to send the functions to the function status module 2282. Typically, the function determination module 2280 selects only one function as the result of each vector. The function status module 2282 is configured to, upon receiving the determined function, evaluate whether the function is healthy based on the vector. When a fault occurs, the functional status module 2282 is configured to send the fault result to the reporting module 2284. In some embodiments, the function status module 2282 may also determine that the status of the function is not related to the operating function of the platform, but is related to the product itself, and may then send the results to the product department to process the feedback accordingly.

The reporting module 2284 is configured to, upon receiving the fault status from the function status module 2282, retrieve the functions determined by the function determination module 2280 and send the functions and the status of the functions to the administrator communicator 230. In certain embodiments. The function status module 2282 may send the function and the status of the function directly to the administrator communicator 230 and a separate reporting module 2284 is not required.

The administrator communicator 230 is configured to, upon receiving the function and fault conditions, match the function to an administrator based on administrator-function relationships (or function owner knowledge) stored in the database 232, prepare an alarm based on determining the function and fault conditions, match the function to the administrator of the function, and send the alarm to the administrator. The administrator of this function, upon receiving the alarm, repairs the fault while it is in charge. If the fault is not associated with the administrator or the administrator is not responsible, the administrator will send a response back to the administrator communicator 228. The administrator communicator 228 then stores the response to the database 232. The AI processor 224 and the AI classifier 228 may then be retrained using the updated database 232. In some embodiments, the response from the administrator may also be stored directly to the database 232.

The database 232 includes data used to train and use the AI processor 224 and the AI classifier 228. Referring to fig. 8, the database 232 includes feedback content 2320, feedback context 2322, user profile 2324, feedback vector 2326, feedback function 2328, feedback status 2330, function administrator list 2332, administrator response 2334, and feedback repair 2336. The feedback content 2320 includes high quality historical feedback from the user and new feedback to be processed by the platform monitor 220. The feedback context 2322 stores the context in which the feedback was submitted. User profile 2324 includes a user's profile, which may include all users registered with the e-commerce website, or only users who have submitted feedback. The feedback vector 2326 stores feedback vectors of historical feedback, which may be used to train the AI processor 224. Feedback functions 2328 include groups of functions corresponding to historical feedback. Feedback status 2330 includes a set of statuses that correspond to historical feedback. The function administrator list 2332 lists the correspondence between functions and administrators responsible for the functions. Administrator responses 2334 include responses from administrators regarding false alarms generated by administrator communicator 230. Failover 2336 stores methods to resolve the failure, if available. The data in the database is indexed by an identification, such as a user ID of a user registration or a session ID of an action, and the data can be retrieved using these identifications.

In certain embodiments, database 232 includes data used to train platform monitor 220. In certain embodiments, database 232 also includes data used during operation of platform monitor 220. In certain embodiments, the database 232 may not include all of the components listed above, and some of the data listed in the database 232 may be stored in other servers or computing devices, accessible by the platform monitor 220 during operation. For example, the function administrator list 2332 may be stored in another device accessible to the administrator communicator 230; feedback fixes 2336 may be stored by a respective administrator. In some embodiments, database 232 may include other documents needed for the operation of platform monitor 220. In certain embodiments, once the new feedback is analyzed, the corresponding feature vectors generated by the AI processor 224, the functions and states determined by the AI classifier 228, and optionally the administrator responses, are stored in the database 232 to update the database 232. In certain embodiments, the AI processor 224 and the AI classifier 228 are retrained periodically or the AI processor 224 and the AI classifier 228 are retrained each time a false alarm is generated.

Fig. 9 schematically illustrates a method of training an AI processor and AI classifier according to certain embodiments of the present disclosure. In certain embodiments, the method is implemented by the server computing device 210 shown in fig. 2. In certain embodiments, the AI processor and the AI classifier are trained independently, with the output of the AI processor, the feature vector of the historical feedback, being used as the input to the AI classifier. It should be particularly noted that, unless otherwise stated in this disclosure, the steps of the method may be arranged in a different order and are therefore not limited to the order shown in fig. 9.

At process 902, high quality historical user feedback is provided. Raw text, images, audio, video are collected from each day's operations of the e-commerce platform. All data is stored in an internal database, such as database 232, and may be used to train the AI algorithm. In certain embodiments, process 902 is performed by feedback extraction module 222.

At process 904, the raw feedback data is cleaned to remove noise. In some embodiments, the process 904 is performed using a content preprocessor 2240 or any other independent module (which may be an AI model). The original text, images, audio and video are noisy and some data is not relevant to the fault. Such as non-informational characters typed by the user, meaningless words uploaded, or low quality images/video. In certain embodiments, one or more AI models are trained to recognize noise patterns and remove noise accordingly. In one embodiment, natural language processing is used to match and remove noisy text. In one embodiment, images with complex noise backgrounds are removed, as these images are typically not screenshots that reflect app/website errors.

Furthermore, the key information is usually only a small part of the original content, so information extraction is crucial. Video is divided into audio and images. The audio is unloaded as text [4 ]. Text [5] in the image is extracted and the image background is left. This is because most users submit images as screenshots when an application fails.

The cleaned images and text are then stored in a database along with contextual features (e.g., timestamps, location of feedback submission, and submission channel) and user profile features. Together, they form the training data for the platform monitor 220, and in particular the AI processor 224 and the AI classifier 228.

At process 906, the data is tagged. In addition to feedback, the label of the feedback is also essential for the success of AI [6 ]. The tag establishes a link between the feedback and its use-a failure of the online retail platform.

Referring back to FIG. 1, we need to map failures to a specific team. Hundreds of categories are defined, each representing a unique functional problem. The quality of the label affects the accuracy of the AI based on its training, so it must be of high quality. In some embodiments, the tags are obtained from historical platform fault reports from a group of professionals who have many years of experience in manually tagging user feedback.

In some embodiments, the functionality of the online retail platform is defined in a tree structure (tree structure), wherein a number of layer 1 functional modules are divided into a plurality of layer 2 modules, each layer 2 module is further divided into a plurality of layer 3 modules, and so on. There are hundreds of modules that serve as leaves of a tree. Through tagging structure and training, the AI can learn how to tag the input feedback to a leaf.

In certain embodiments, the AI processor is fixed for feedback, adjusting parameters of the AI classifier to refine the AI classifier according to the quality of the results, where the results may be a percentage of correct alarms generated by the AI classifier; in other embodiments, the AI classifier is fixed and the parameters of the AI processor are adjusted to refine the AI processor according to the quality of the results, which may be the percentage of correct alarms generated by the AI classifier. In some embodiments, the above-described methods may also be used to select an appropriate AI model for a portion of the platform monitor 220. In other words, by fixing the AI classifier and changing one of the AI processors, an appropriate AI processor model can be selected; and by fixing the AI processor and changing one of the AI classifiers, a suitable AI classifier model can be selected.

FIG. 10 schematically illustrates the structure of functions to be monitored, according to certain embodiments of the present disclosure. As shown in fig. 10, the functional structure of the monitored platform 1000 includes three layers. Platform 1000 may be an e-commerce platform. The layer 1 modules of the platform 1000 include a website 1010 of the e-commerce platform, an App 1030, such as a smart phone application, and an external 1050, such as a third party service, that communicates with the e-commerce platform. The layer 1 module here is the submission conduit for platform feedback. The layer 1 module website 1010 includes three layer 2 modules, namely a product page 1012, a shopping cart 1014, and a payment 1016. The user may sequentially access the three layer 2 modules of the website 1010 and show the three layer 2 modules in different web pages. For example, the user can browse the products page 1012 and find products of interest to him, add products to the shopping cart through the shopping cart function 1014, view the shopping cart, and make payments through the payment function 1016. The layer 2 module product page 1012 further includes four layer 3 modules including a product description 1012A, a product search 1012B, a product recommendation module 1012C, and other related modules. These three layer 3 functions are directly related to the product and can be displayed in the same web page.

In some embodiments, the tag class is defined manually, so it is not perfect. Some classes are not defined and new classes are emerging when developing new functionality for a platform.

For the former, we collect feedback that is not classified into any predefined categories as "unknown feedback". There are usually multiple unknown classes, so we must further segment the unknown feedback into subgroups [6] via unsupervised machine learning. Topic modeling (natural language processing techniques [7]) is used to extract topic information and perform human intervention to define those classes that are undefined.

For the latter, we follow a similar procedure. The only difference is that the new category usually does not have much feedback. Its feedback is "left behind" after the previous step, independent of any defined categories. Finally, we map newly published functions and match them with legacy functions.

After tagging the data, the original content of the cleaned feedback is converted to a numerical value at process 908 through process 912. These values are integrated with the context characteristics and the user profile characteristics to form a matrix (feedback vector) corresponding to the feedback. The process 908 through the process 912 are executed to train the AI processor 224.

At process 908, text processing module 2242 receives the text for feedback and, in response, processes the text to obtain features or values for the text. In some embodiments, text processing module 2242 splits each text into a sequence of words and then converts each word into a number, representing word occurrence, word co-occurrence, part of speech, named entities, sentence syntax structure, and word semantics. These numbers are the dimensions of the feature vector for each feedback, respectively. In some embodiments, the characterization of multiple words for feedback results in a large matrix with columns as the above-mentioned features and rows as the feedback. The matrix is called a text matrix.

At process 910, the image processing module 2246 receives the fed back image and converts the image to a numerical value. Specifically, image processing module 2246 separates the image into text (image text) and a background image (image without text) extracted from the image, processes the image text and the background image, respectively, and integrates the results to obtain a numerical value, and adds the value as a new dimension of the text matrix. In some embodiments, audio and video are converted to text and images and similarly processed to obtain their respective values. These values are added to the text matrix as a new dimension of the vector, where each row corresponds to one feedback and is treated as a vector of that feedback. In some embodiments, when the feedback includes only text, process 908 is sufficient and process 910 is not necessary.

At process 912, the vector generator 226 extracts information from the context of the feedback and the user profile, converts the information to values, and adds the values to the text matrix to form a feedback matrix (or named feature matrix). Referring back to fig. 6, a feedback matrix is shown in accordance with certain embodiments of the present disclosure. The rows of the matrix are vectors of feedback. Each row (vector of each feedback) includes text features (word features, syntactic features, semantic features), image text features, background image features, context features, and user profile features, which are represented by values, the values in a row being the dimensions of the vector corresponding to the feedback. In some embodiments, the feedback in a row may not have a corresponding characteristic for each dimension, and dimensions lacking in the feedback are assigned a value of 0.

Certain dimensions of the feature matrix are obtained by running the AI processor, and after obtaining the feedback matrix, the AI classifier 228 is trained at process 914 using the matrix and corresponding functional label (or fault label) as inputs. Fig. 11 schematically illustrates an integrated structure according to certain embodiments of the present disclosure. Referring to FIG. 11, various machine learning classification models are applied, followed by an integration mechanism [6 ].

Given the unbalanced distribution of feedback tags (some classes of feedback are much less than others), the feedback data is resampled via bootstraps [6] to make the tags more evenly distributed.

A Gradient enhancement tree classifier [11] is applied as the main classifier (e.g., classification model 1 in fig. 11). The data is a combination of a class-value type and a continuous-value type, and is unbalanced. Gradient enhancement trees are known to be successful in data characterization [11 ]. Furthermore, to reduce bias by using a single model, a hierarchical integration mechanism is applied. The integration uses its companion model to synthesize a gradient-enhanced tree, such as a random forest, logistic regression with penalty terms [6 ]. Each classifier outputs a prediction tag. The classifier, named layer 1 classifier, includes a classification model 1, a classification model 2. Each classification model corresponds to a function of the platform. When layer 1 of the feature matrix is used as input for each model, each model gives a result of whether the function corresponding to the model is referred to in the feedback. The results of all classification models 1 through N are combined to form the feature matrix level 2. For example, each result from one of the classification models is in the form of a binary decision, and the binary decisions from the classification models 1 through N are respectively defined as dimensions of a vector. Here, the vector is a feature matrix layer 2 when one feedback is processed, or one row of a feature matrix layer 2 when a plurality of feedbacks are processed. The feature matrix layer 2 information is integrated into an ensemble classifier that provides the result of a function whether the ith label is feedback-related or feedback-unrelated. In some embodiments, the output of whether the ith label is related to feedback is in the form of a binary decision.

One uniqueness of this step is that the layer 1 context and user features are reused in layer 2. This is because during training, it is observed that although these features have strong signals, they are much smaller in number than the text/image features. So when training the layer 1 classifier, its signal is buried in a large number of text/image features. After layer 1, the text/image features are reduced to class labels, so in the layer 2 classifier, the text/image features are not much, and if we use the context/user features as the input of layer 2, the ability can be fully utilized. In some embodiments, another gradient enhanced tree classifier is used as the integrated classifier.

The process shown in FIG. 11 is for predicting individual tags. In other words, it determines the relationship between the matrix (or vectors in the matrix) and one particular function. This process may be repeated for each tag or function. The results of all these tags are then integrated to make the final decision of which tag to give feedback, as shown in fig. 12. In other words, the process in FIG. 12 finds a function from among a plurality of functions, where the function is related to or referred to in the feedback, when determining whether the feedback is related to each function.

In some embodiments, the models are trained one-by-one using a one-to-many method [6 ]. This is because there are hundreds of classes, and if we use one model for all classes, there will be hundreds of variables for optimization, and it is difficult for the computer to find the best model. One-to-many means we treat one category as positive and all other categories as negative. In other words, we train hundreds of layer 2 classifiers, each predicting whether the feedback belongs to a certain class. Since all of these classifiers are trained separately, they may be divergent from each other, and multiple classifiers may consider the feedback to belong to their corresponding classes, but in practice the feedback may belong to only one of them. Therefore, there is a need to integrate these hundreds of opinions and find a way to get the best consensus based on them. FIG. 12 schematically illustrates a method of integrating all one-to-many classifiers, according to some embodiments of the present disclosure. Note that it is similar to fig. 11, but serves a different purpose. Fig. 11 is predicting individual tags, while fig. 12 is integrating all of these tags to make the final decision of which tag to give feedback. The tags represent specific failures owned by a particular team, so the tags are used to map to a particular team and send failure notifications to the team.

In some embodiments, the machine learning model is retrained and updated periodically to learn the latest classification patterns in the data. In some embodiments, an online training mechanism is enabled. Once the AI makes a mistake, e.g., a false alarm is sent, an error message will be immediately sent back to the AI model for on-line retraining. The system is therefore self-perfecting in real time. The classification results will be sent to the function owners, including developers, product experts, and analysts, so that certain measures can be taken to fix the detected problem.

After the AI processor 224 and AI classifier 228 are trained, the platform monitor 220 may be used to monitor the health of the e-commerce platform. FIG. 13 schematically illustrates a method of checking the health of an e-commerce platform using a platform monitor 220 based on feedback from a user. In certain embodiments, the method is implemented by the server computing device 210 shown in FIG. 2. It should be particularly noted that, unless otherwise stated in this disclosure, the steps of the method may be arranged in a different order, and thus are not limited to the order as shown in fig. 13.

As shown in fig. 13, at process 1302, the feedback extraction module 222 extracts the content of the feedback. The content may include at least one of text, voice, image, and video.

At process 1304, the content preprocessor 2240 cleanses the content of the feedback.

After cleansing the content of the feedback, at process 1306, the AI processor 224 processes the cleansed content to obtain a text vector and sends the feature vector to the vector generator 226. The dimensions of the text vector are values corresponding to text in the content, image text in the content, and a background image.

The vector generator 226 then adds the context of the feedback and the user profile to the text vector as a new dimension to form a feature vector at process 1308, and sends the feature vector to the AI classifier 228. The new dimensions include time of submission, location of submission, channel of submission, user attributes, user history, and user preferences.

At process 1310, in response to receiving the feature vector, the classifier 228 processes the feature vector to obtain a fault corresponding to the feature vector and sends fault information to the administrator communicator 230. Specifically, the classifier 228 determines whether the feature vector is function-related, whether the state of the function represented by the feature vector is normal or abnormal/faulty, and which function the feature vector represents.

At process 1312, in response to receiving the fault information, the administrator communicator 230 matches the fault with a particular administrator or team responsible for the fault, prepares an alarm, and sends the alarm to the administrator.

After receiving the alarm, the administrator determines whether the fault is one for which he is responsible. If he is responsible for the failure, he will fix the failure. If not, the administrator sends a response to the administrator communicator 230 of the platform monitor 220.

At process 1314, the administrator communicator 230 receives a response from the administrator that includes information that the feedback or fault is not responsible for the administrator.

At process 1316, the administrator communicator 230 stores the error information from the administrator to the database 232 and retrains the AI classifier 228 using the updated database 232.

In summary, a platform monitor according to certain embodiments of the present disclosure is a self-contained, self-perfecting system. By utilizing the user's feedback (content, context, and profile) and integrating it with knowledge about the platform (corporate team structure, application functionality structure, and their correspondences), the system can timely, accurately, and automatically detect functional problems via the utilization of the capabilities of artificial intelligence including natural language processing, computer vision, and machine learning.

The content, context and profile of the feedback is converted into a vector with a large number of dimensions, which makes the final fault decision accurate.

The number of vector dimensions is easy to expand, and the function category structure is easy to expand, so that the newly added information or functions are convenient to merge.

In the training of AI, the context and profile of feedback are used under the layer 1 functional model and the layer 2 functional model, so that the influence of the context and profile information can be effectively considered and can not be submerged by the influence of the feedback content.

The foregoing description of the exemplary embodiments of the present disclosure has been presented for the purposes of illustration and description only and is not intended to be exhaustive or to limit the disclosure to the precise forms disclosed. Many modifications and variations are possible in light of the above teaching.

The embodiments were chosen and described in order to explain the principles of the disclosure and its practical application to enable others skilled in the art to utilize the disclosure and various embodiments and with various modifications as are suited to the particular use contemplated. Alternative embodiments will become apparent to those skilled in the art to which the present disclosure pertains without departing from its spirit and scope. Accordingly, the scope of the present disclosure is defined by the appended claims rather than the foregoing description and the exemplary embodiments described herein.

Reference documents:

1.Tomas Mikolov,Ilya Sutskever et al,Distributed Representations ofWords and Phrases and their Compositionality

2.Quoc Le,Tomas Mikolov,Distributed Representations of Sentences andDocuments

3.Yoon Kim,Convolutional Neural Networks for Sentence Classification

4.Pieraccini,Roberto.The Voice in the Machine.Building Computers ThatUnderstand Speech.The MIT Press

5.https://github.com/tesseract-ocr/

6.Trevor Hastie,Robert Tibshirani,and Jerome H.Friedman,The elementsof statistical learning,2001,Springer

7.Blei,David,Probabilistic Topic Models,Communications of the ACM,2012,55(4):77–84

8.Christopher Manning,Hinrich Schutze,Foundations of statisticalnatural language processing,1999,The MIT Press

9.Mikolov,Tomas；et al.Efficient Estimation of Word Representations inVector Space,arXiv:1301.3781

10.LeCun,Yann.LeNet-5,convolutional neural networks.Retrieved 16November 2013

11.https://github.com/dmlc/xgboost.

Claims

1. a method for monitoring an e-commerce platform, the method comprising:

receiving, by a computing device, feedback submitted by a user through an e-commerce platform;

generating, by a feedback processor of the computing device, a vector based on content of the feedback, context of the feedback, and a user profile, wherein the content comprises at least one of text, speech, images, and video, wherein the context comprises at least one of a time at which the feedback was submitted, a location at which the feedback was submitted, and a submission channel of the feedback, wherein the user profile comprises at least one of attributes of the user, a purchase history of the user, and preferences of the user using the e-commerce platform; and

classifying, by a classifier of the computing device, the vector to obtain a function of the e-commerce platform and a status of the function corresponding to the feedback and to prepare an alert when the status is a fault,

wherein the vector includes a predetermined number of dimensions and each of text, speech, images, video, time at which the feedback was submitted, location at which the feedback was submitted, channel of submission of the feedback, attributes of the user, purchase history of the user, and preferences of the user corresponds to at least one dimension of the vector.

2. The method of claim 1, wherein the feedback processor and the classifier are performed using at least one artificial intelligence model.

3. The method of claim 2, wherein generating the vector comprises: processing the content using the feedback processor to obtain content dimensions for the vectors corresponding to text, speech, images, and video.

4. The method of claim 3, further comprising: cleansing the content to obtain content dimensions for the vector prior to processing the content.

5. The method of claim 3, further comprising: separating the image into a text of the image and a background image, processing the text of the image to obtain an image text result, processing the background image to obtain a background image result, and integrating the image text result and the background image result to obtain a content dimension of the vector corresponding to the image.

6. The method of claim 2, further comprising: sending the alert to an administrator of the e-commerce platform responsible for the function, receiving an instruction from the administrator corresponding to the alert when the alert is false, and retraining the feedback processor and the classifier using the instruction.

7. The method of claim 2, wherein the classifier is trained using a plurality of historical feedback and functional category structures, the functional category structures including layer 1 categories, the layer 1 categories including websites of an e-commerce platform, applications of an e-commerce platform, and external links of an e-commerce platform.

8. The method of claim 7, wherein the layer 1 category of websites includes a layer 2 category, the layer 2 category including product pages, shopping carts, and payments.

9. The method of claim 8, wherein the layer 2 category of the product page includes a layer 3 category, the layer 3 category including product description, product search, and product recommendation.

10. The method of claim 7, wherein the classifier includes a plurality of classification models, each classification model providing a candidate function based on each historical feedback, an integration model using the candidate functions provided by the classification models to determine a function corresponding to each feedback.

11. A system for monitoring an e-commerce platform, the system comprising a computing device comprising a processor and a storage device storing computer-executable code, wherein the computer-executable code is configured to, when executed at the processor:

receiving feedback submitted by a user through an e-commerce platform;

generating a vector based on content of the feedback, context of the feedback, and a user profile, wherein the content comprises at least one of text, speech, images, and video, wherein the context comprises at least one of a time at which the feedback was submitted, a location at which the feedback was submitted, and a channel of submission of the feedback, wherein the user profile comprises at least one of attributes of the user, a purchase history of the user, and preferences of the user using the e-commerce platform; and

classifying the vector to obtain a function of the e-commerce platform and a status of the function corresponding to the feedback and to prepare an alarm when the status is a failure,

12. The system of claim 11, wherein the computer-executable code comprises a feedback processor for generating the vector and a classifier for classifying the vector, the feedback processor and the classifier comprising an artificial intelligence model.

13. The system of claim 12, wherein the computer-executable code is configured to generate the vector by: the content is cleaned and processed using the feedback processor to obtain content dimensions for the vectors corresponding to text, speech, images, and video.

14. The system of claim 13, wherein the computer-executable code is further configured to: separating the image into a text of the image and a background image, processing the text of the image to obtain an image text result, processing the background image to obtain a background image result, and integrating the image text result and the background image result to obtain a content dimension of the vector corresponding to the image.

15. The system of claim 12, wherein the computer-executable code is further configured to: sending the alert to an administrator of the e-commerce platform responsible for the function, receiving an instruction from the administrator corresponding to the alert when the alert is false, and retraining the feedback processor and the classifier using the instruction.

16. The system of claim 11, wherein the classifier is trained using a plurality of historical feedback and a functional class structure comprising:

a layer 1 category comprising websites of the e-commerce platform, applications of the e-commerce platform and external links of the e-commerce platform,

the web sites of the layer 1 category include a layer 2 category, the layer 2 category including product pages, shopping carts, and payments, an

The product page of the layer 2 category comprises a layer 3 category, and the layer 3 category comprises product description, product search and product recommendation.

17. The system of claim 16, wherein the classifier includes a plurality of classification models, each classification model providing a candidate function based on each historical feedback, the integration model using the candidate functions provided by the classification models to determine a function corresponding to each feedback.

18. A non-transitory computer-readable medium storing computer-executable code, wherein the computer-executable code is configured to, when executed at a processor of a computing device:

receiving feedback submitted by a user through an e-commerce platform;

19. The non-transitory computer-readable medium of claim 18, wherein the computer-executable code comprises a feedback processor to generate the vector and a classifier to classify the vector, the feedback processor and the classifier comprising an artificial intelligence model.

20. The non-transitory computer-readable medium of claim 18, wherein the computer-executable code is configured to process the image by: separating the image into text of the image and a background image, processing the text of the image to obtain an image text result, processing the background image to obtain a background image result, and integrating the image text result and the background image result to obtain the dimension of the vector corresponding to the image.